██╗ ██████╗ ██████╗ ███████╗██╗ ██╗███████╗████████╗███████╗███╗ ███╗
██║ ██╔══██╗██╔══██╗ ██╔════╝╚██╗ ██╔╝██╔════╝╚══██╔══╝██╔════╝████╗ ████║
██║ ██████╔╝██████╔╝ ███████╗ ╚████╔╝ ███████╗ ██║ █████╗ ██╔████╔██║
██║ ██╔═══╝ ██╔══██╗ ╚════██║ ╚██╔╝ ╚════██║ ██║ ██╔══╝ ██║╚██╔╝██║
███████╗██║ ██║ ██║ ███████║ ██║ ███████║ ██║ ███████╗██║ ╚═╝ ██║
╚══════╝╚═╝ ╚═╝ ╚═╝ ╚══════╝ ╚═╝ ╚══════╝ ╚═╝ ╚══════╝╚═╝ ╚═╝
A full-stack web-based license plate detection and recognition system built for traffic surveillance. Upload video footage from multiple CCTV sources, detect license plates in real-time, extract text via OCR, and search through the complete detection history — all from a sleek dashboard.
The main dashboard is the first screen you see when the app launches. The top navigation bar contains 4 tabs:
- CCTV 1 — Dedicated channel for Camera 1 footage
- CCTV 2 — Dedicated channel for Camera 2 footage
- Video — For uploading and processing pre-recorded video files
- History — Full searchable detection log with record count badge
The left panel shows the video file list for the selected source, an upload zone (drag & drop or browse), and action buttons — Process Selected, Stop, and Refresh. The right panel shows the Detected Plates feed. The bottom section shows the Live View canvas where the annotated video stream renders in real time.
Once a video is selected and Process Selected is clicked, the system connects via WebSocket and begins processing the video frame by frame. The live view updates in real time with green bounding boxes drawn around every detected license plate.
The bottom stats bar updates continuously:
- 600 Frames processed
- 356 Total detections made
- 356 Unique plates identified
- 6 FPS processing speed
A red progress bar fills up showing how much of the video has been processed. The status pill in the header switches from
IdletoProcessing(green dot) while active.
The right-side Detected Plates panel updates in real time as each plate is recognized. Every detection card shows:
- Cropped plate thumbnail — the actual image extracted from the video frame
- Plate text — recognized via EasyOCR (e.g.
LP14 LJA,NL64 OGX)- Confidence % — YOLO detection confidence score
- Source — which channel (CCTV 1 / CCTV 2 / VIDEO)
- Frame number — which frame the plate was detected in
- Timestamp — exact time of detection
A red confidence bar at the bottom of each card gives a visual indicator of detection quality. The badge in the header shows the total count — 61 detections in this view.
The Detection History tab shows a complete table of every plate ever detected across all sessions. The History tab badge shows 442 total records accumulated.
Each row contains:
- # — Record index
- Plate — Recognized plate text (bold white)
- Confidence — Detection confidence in red
- Source — Source badge (VIDEO / CCTV1 / CCTV2)
- Frame — Frame number where the plate was found
- Image — Thumbnail of the cropped plate
- Timestamp — Date and time (
2026-05-23 12:30:07)
The Search feature allows instant filtering of detection history. Searching for
R-183-JFreturns exactly 1 result — detected at 84.4% confidence from the VIDEO source at frame 12 on2026-05-23 12:23:45.The search bar supports partial matches. The Source dropdown allows filtering by
All Sources,CCTV 1,CCTV 2, orVideo. The cropped plate thumbnail is shown for visual verification.
| Feature | Description |
|---|---|
| 🎯 YOLO26 Detection | Latest Ultralytics model — NMS-free, 43% faster inference than YOLO11 |
| 🔤 EasyOCR Recognition | GPU-accelerated OCR engine for plate text extraction |
| 📷 Multi-Source Support | CCTV 1, CCTV 2, and Video tabs — independent file management per source |
| 🎬 Real-Time Live View | Annotated WebSocket video stream rendered in the browser |
| 🖼️ Auto Plate Crop & Save | Every detected plate cropped and saved as .jpg in output/plates/ |
| 💾 CSV Detection Log | All detections saved with plate text, confidence, source, frame, timestamp |
| 🔍 Search & Filter | Search by plate number, filter by source channel instantly |
| 📊 Live Stats Dashboard | Real-time Frames, Detections, Unique Plates, and FPS counters |
| 🌐 Web-Based UI | Dark red themed dashboard — no extra frontend framework required |
| ⚡ GPU Acceleration | CUDA-powered inference on NVIDIA GPU |
| 🔄 Duplicate Filtering | Same plate seen multiple times per video is deduplicated |
| Component | Technology |
|---|---|
| Object Detection | YOLO26 (Ultralytics 8.4+) |
| OCR Engine | EasyOCR 1.7 |
| Backend | FastAPI + WebSocket |
| Frontend | HTML5 / CSS3 / Vanilla JS |
| Video Processing | OpenCV 4.10 |
| Data Storage | CSV (no database needed) |
| GPU Support | CUDA 12.1 (NVIDIA) |
| Model Training | Roboflow Dataset + YOLO26 Transfer Learning |
License-Plate-Detection-and-Recognition/
│
├── 📁 models/
│ └── best.pt # Trained YOLO26 model
│
├── 📁 uploads/ # Uploaded input videos
│
├── 📁 output/
│ └── 📁 plates/ # Cropped plate images
│
├── 📁 data/
│ └── detections.csv # Complete detection log
│
├── 📁 templates/
│ └── index.html # Web dashboard UI
│
├── 📁 screenshots/ # Project screenshots
│
├── app.py # FastAPI backend (main app)
├── train.py # YOLO26 training script
├── download_dataset.py # Roboflow dataset downloader
├── requirements.txt # Python dependencies
└── README.md
- Python 3.10+
- Windows or Linux
- NVIDIA GPU (recommended) with CUDA 12.1+
- Video files (MP4, AVI, MOV, MKV)
git clone https://github.com/naimurhamim/License-Plate-Detection-and-Recognition.git
cd License-Plate-Detection-and-Recognitionpython -m venv lpr_env
# Git Bash (Windows)
source lpr_env/Scripts/activate
# PowerShell (Windows)
lpr_env\Scripts\activate# Check your CUDA version first
nvidia-smi
# CUDA 12.1+
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
# CUDA 11.8
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118pip install -r requirements.txtpython app.pyOpen browser: http://localhost:8000 🚀
python download_dataset.pyDataset: License Plate Recognition — Roboflow Universe
| Stat | Value |
|---|---|
| Total Images | 24,242 |
| mAP@50 | 97.2% |
| Precision | 100% |
| Recall | 100% |
| Classes | 1 (License_Plate) |
python train.pyBest model saved at:
runs/lpr_yolo26/weights/best.ptCopy tomodels/best.ptto use in the app.
Video Upload (MP4 / AVI / MOV / MKV)
↓
Frame Extraction — every 3rd frame for speed
↓
YOLO26 Detection → Bounding box around license plate
↓
Plate Region Crop → saved as .jpg in output/plates/
↓
EasyOCR → Extract plate text from cropped image
↓
Duplicate Filter → skip if same plate already seen this video
↓
CSV Append → plate_text, confidence, source, frame_num, timestamp
↓
WebSocket → Send annotated frame + detection data to browser
↓
Browser UI → Live view canvas + Detection panel updates
| Parameter | Default | Description |
|---|---|---|
| Detection confidence | 0.4 |
Minimum YOLO confidence threshold |
| Frame skip | Every 3rd | Process every 3rd frame for speed |
| OCR language | en |
EasyOCR language |
| GPU | True |
Set gpu=False in app.py for CPU only |
| Server port | 8000 |
Change in last line of app.py |
Tested on NVIDIA GeForce RTX 3050 Ti Laptop GPU (4GB VRAM)
| Metric | Value |
|---|---|
| Processing Speed | ~6 FPS |
| YOLO26 mAP@50 | 97.2% |
| Unique plates per video | 300–500+ |
| Detection confidence avg | 70–85% |
| Supported formats | MP4, AVI, MOV, MKV |
ultralytics>=8.3.0
easyocr>=1.7.0
fastapi>=0.115.0
uvicorn[standard]>=0.30.0
websockets>=12.0
python-multipart>=0.0.9
aiofiles>=23.2.1
jinja2>=3.1.0
opencv-python>=4.10.0
numpy>=1.26.0
python-dotenv>=1.0.0
Pillow>=10.0.0
roboflow>=1.3.0
⚠️ Install PyTorch before runningpip install -r requirements.txt
- Fork the repo
- Create your feature branch:
git checkout -b feature/AmazingFeature - Commit:
git commit -m 'Add AmazingFeature' - Push:
git push origin feature/AmazingFeature - Open a Pull Request
Licensed under the MIT License — see LICENSE for details.
